Face Sketch Synthesis Based On Machine Learning

Sketch recognition is a new branch of face recognition. In applications of case-solving and suspect-searching, to deal with the problem that the photo of a suspect is unavailable in most case, a simulated sketch is generated by cooperation of artists and eyewitnesses as a substitute, so that the content-based image retrieval for identification is conducted in the existing photo database. Due to large differences between a photo-sketch pair caused by different generation mechanism and information expression manner, it is impossible to recognize by simply matching the photo-sketch pair. As a result, how to reduce the difference becomes important for sketch-photo recognition, and sketch synthesis becomes the key technique of recognition.Sketch synthesis algorithm based on Embedded Hidden Markov Models (E-HMM) is proposed firstly, in which the pseudo-sketch of the given photo is synthesized with the nonlinear mapping between a photo-sketch pair which is learnt by E-HMM. Unfortunately, the performance depends on the Viterbi decoding sequence which is easily influenced by noise. Therefore, the E-HMM Inversion (E-HMMI) algorithm, which synthesizes the pseudo-sketch with the joint model without Viterbi decoding, is introduced to improve the quality of the pseudo-sketch. Experimental results show that isolated noise can be removed by E-HMMI in large degree to smooth the pseudo-sketch.Furthermore, aiming at the disadvantage that the complicated relationship cannot be learnt with single photo-sketch pair, sketch synthesis based on E-HMM and Selective Ensemble (SE) is proposed by adopting SE to improve the generalization ability of the system. Experimental results show that the novel algorithm can achieve good results in small sample set. Based on the above algorithm and the fact that local information can represent face better than holistic information, a sophisticate sketch synthesis algorithm is presented based on machine learning and local information.Finally, by combining the full reference and the no reference image quality assessment, which corresponds to the similarity between the pseudo-sketch and the original photo and the quality of the sketch itself, a sketch quality assessment metric is proposed. Experimental results illustrate the good consistency between the proposed objective metric and subject assessment.